Tutorials Of Scene Segmentation Model Server

Start A Scene Segmentation Server

It’s very quick to start a scene segmentation server. Main code are showed below

Scene Segmentation Server Code Snappit

The executable binary file was built in $PROJECT_ROOT/_bin/bisenetv2_segmentation_server.out Simply run

cd $PROJECT_ROOT/_bin
./bisenetv2_segmentation_server.out ../conf/server/scene_segmentation/bisenetv2/bisenetv2_server_config.ini

When server successfully start on http:://localhost:8091 you’re supposed to see worker_nums workers were called up and occupied your GPU resources. By default 4 model workers will be created you may enlarge it if you have enough GPU memory.

Python Client Example

Local python client test is similiar with mobilenetv2 classification server you may read toturials_of_classfication_model_server.md for details.

To use test python client you may run

cd $PROJECT_ROOT/scripts
export PYTHONPATH=$PWD:$PYTHONPATH
python server/test_server.py --server bisenetv2 --mode single

Unique Tips For Scene Segmentation Model Python Client

Scene Segmentation model’s output is a class map with the same image size of origin input image. Each pixel was assigned with a unique class label. Server’s response is a json like

resp = {
    'req_id': '',
    'code': 1,
    'msg': 'success',
    'data': {
        'segment_result': base64_image_content
    }
}

segmentation_result contains the model’s output encoded with base64. If you want to save the model’s output info local file you may do

with open(src_image_path, 'rb') as f:
    image_data = f.read()
    base64_data = base64.b64encode(image_data)

    post_data = {
        'img_data': base64_data.decode(),
        'req_id': 'demo',
    }
    resp = requests.post(url=url, data=json.dumps(post_data))
    output = json.loads(resp.text)['data']['segment_result']
    out_f = open('result.png', 'wb')
    out_f.write(base64.b64decode(output))
    out_f.close()

Scene Segmentation Model’s Visualization Result

BisenetV2 Model

BisenetV2 :fire: model was designed for fast scene segmentation task. You may refer to repo https://github.com/MaybeShewill-CV/bisenetv2-tensorflow for details about training details.

Network’s main structure is Bisenetv2 Network Architecture

Server's Input Image

Server's Output Image